把 Jev 接进驾驶模拟器跑决策循环

grokked @grokkedd原推发布于 2026.09.20

在 X 上打开原推

Someone already tried wiring Jev into a driving simulator, and it's a genuinely interesting stress test Jev, TypeSafe's decision-only model, is showing up in early experiments with autonomous driving. According to the person running it, a full decision loop for a simulated car went together in under an hour on the new model: turns, lane changes, intersections, merging onto a highway. For every action, the model doesn't return text, it returns typed probabilities across the possible decisions along with a confidence score. That's the actual difference from a regular chat model here. You feed in system state, and you get back a set of allowed actions with probabilities attached, not a paragraph you have to parse into a decision. Worth being clear about the limits before this gets overhyped: the demos circulating so far, including a similar real-time Doom control clip, feed the model clean structured text state, positions, distances, angles, not raw camera input. Jev has no vision component. Real self-driving needs on-device decisions in the 10-20ms range with actual sensor fusion, and Jev's own benchmarks put it at 70-500ms for a single decision. A simulated environment handing over perfect JSON state is a much easier problem than reading the road. Still, if this kind of typed-decision approach scales past toy simulations, it could meaningfully lower the barrier to prototyping decision layers for autonomous systems, even if the perception stack stays somebody else's problem entirely.

这条是引用着另一条写的：x.com/grokkedd/status/2100151292404302270